Schematic of different convolution operations: (a) ordinary convolution, (b) residual convolution, and (c) residual grouped convolution modules. (IMAGE)
Caption
Through a comprehensive evaluation of model complexity and number of parameters, it was determined that the overall performance of the proposed model is the best when eight group convolutions are used. Therefore, in the proposed model, the high-dimensional convolution was grouped into eight identical low-dimensional convolutions, which effectively solved the aforementioned problems. When the input 64-channel image undergoes a 3 × 3 convolution for a 64-channel image output, there are 36864 parameters in this layer. In the proposed model, the dimensionality of the input data is first reduced and then passed through eight group convolution layers; the number of parameters in each group is 576, and the total number of parameters from the eight groups is 4608. Finally, the output is upsampled to the size of the original data.
Credit
Beijing Zhongke Journal Publising Co. Ltd.
Usage Restrictions
Credit must be given to the creator.
License
CC BY